使用 GitHub 增強基于 AI 的代碼合成的安全性 通過廉價而高效的即時工程實現(xiàn)副駕駛

EnhancingSecurityofAI-BasedCodeSynthesiswithGitHubCopilotviaCheapandEfficientPrompt-Engineering

JakubRes

ir

esj@fit.vut.cz

BrnoUniversityofTechnology,FacultyofInformationTechnologyCzechRepublic

Ale?Smr?ka

smr

cka@fit.vut.cz

BrnoUniversityofTechnology,FacultyofInformationTechnologyCzechRepublic

ABSTRACT

IvanHomoliak

ihomoliak@fit.vut.cz

BrnoUniversityofTechnology,

FacultyofInformationTechnologyCzechRepublic

KamilMalinka

malinka@fit.vut.cz

person*newPerson=(person*)malloc(sizeof(person));newPerson->status=0;

BrnoUniversityofTechnology,FacultyofInformationTechnologyCzechRepublic

MartinPere?íni

iper

esini@fit.vut.cz

BrnoUniversityofTechnology,FacultyofInformationTechnologyCzechRepublic

PetrHanacek

hanacek@fit.vut.cz

BrnoUniversityofTechnology,FacultyofInformationTechnologyCzechRepublic

arXiv:2403.12671v1

[cs.CR]

19

Mar2024

AIassistantsforcodingareontherise.Howeveroneofthereasonsdevelopersandcompaniesavoidharnessingtheirfullpotentialisthequestionablesecurityofthegeneratedcode.Thispaperfirstreviewsthecurrentstate-of-the-artandidentifiesareasforim-provementonthisissue.Then,weproposeasystematicapproachbasedonprompt-alteringmethodstoachievebettercodesecurityof(evenproprietaryblack-box)AI-basedcodegeneratorssuchasGitHubCopilot,whileminimizingthecomplexityoftheapplica-tionfromtheuserpoint-of-view,thecomputationalresources,andoperationalcosts.Insum,weproposeandevaluatethreepromptalteringmethods:(1)scenario-specific,(2)iterative,and(3)generalclause,whilewediscusstheircombination.Contrarytotheauditofcodesecurity,thelattertwooftheproposedmethodsrequirenoexpertknowledgefromtheuser.WeassesstheeffectivenessoftheproposedmethodsontheGitHubCopilotusingtheOpenVPNprojectinrealisticscenarios,andwedemonstratethattheproposedmethodsreducethenumberofinsecuregeneratedcodesamplesbyupto16%andincreasethenumberofsecurecodebyupto8%.SinceourapproachdoesnotrequireaccesstotheinternalsoftheAImodels,itcanbeingeneralappliedtoanyAI-basedcodesynthesizer,notonlyGitHubCopilot.

INTRODUCTION

WiththereleaseofChatGPT[

1

],publicattentionshiftedtowardsAIassistanttools.Theseassistantsareproficientinmanyareas,includingsoftwareengineeringorcoding.TheadventofAIcodingassistantsmeanstransitioningfromintelligentcode-completiontoolstocode-generatingtools.AlthoughtheseAIassistantsarefarfromperfect,intermsofsolvingcodingproblems,arecentmodelAlphaCode2,proposedbyDeepmind,scoredbetterthanover85%ofhumancompetitors[

9

].

AccordingtoLiangetal.[

11

]inthesurveywith410Githubusers’responses,70%ofrespondentswhohadexperienceswithGithubCopilotutilizeitatleastonceinamonthwhile46%utilizetheAIassistantdaily.ThemostfrequentreasonsfordevelopersusingAIassistantswerefewerkeystrokestowritecodeandfastercoding.DuetotherapidlyrisingpopularityofAIassistants,researchersstartedtofocusonstudyingthequalityofthesynthesizedcodeand

Fig.1:ExampleofsecurityissuegeneratedbyAI.Thesce-nariocomesfromthedatasetproposedin[

17

].

waysofimprovingit(see

Sec.5.2

).Whileobservingthevalidityorcorrectness,manystudiesoverlookthecrucialaspectofcode—security.

Inthemotivatingexample,theAIassistantwastaskedwithgeneratingacodesnippettofillagapinthecontextofaCprogram.Itsobjectivewastocreateanewinstanceofthestructure"person"andassignastatusvalueofzerotoit.AlthoughtheAIassistantprovidedareasonablecode(see

Fig.1

),thesnippetcontainCWE-476[

25

](themallocfunctioncouldfailtoallocatememory,thusresultinginaNULLpointerdereference).

Inthisresearch,weaimtostudyvariouswaysofimprovingcodesecuritygeneratedbyanyproprietaryLargeLanguageMod-els(LLMs),andwedemonstrateourapproachonthewell-knownGitHubCopilot[

6

].

Thereexistafewcategoriesforimprovingthecodesynthe-sisofAImodels,suchasoutputoptimization,modelfine-tuning,andpromptengineering,andeachofthemhassomeprosandcons.Inthiswork,wefocusonefficiency,generality,andlowcosts,andthereforepromptengineeringisthemostsuitabletech-niqueforus.Whileliteratureforpromptengineeringismostlygeneral[

14

][

31

][

5

][

4

],wearemorespecificanddeterminefourap-proachestoit,whichwefurtherinvestigate:(1)scenario-specificinformationandwarningproviding,(2)iterativesecurity-specificprompting,(3)generalalignmentshiftingusinginceptionprompt(i.e.,generalclause),(4)cooperativeagentssystem.Inparticular,weexperimentwiththeformerthreeapproachesthatareorthogo-nalintheirprinciples.

Contributions.Thecontributionsofourpaperareasfollows:

WereviewedtheliteratureandidentifiedthreedifferentareasofcodesynthesisimprovementsofLLMs,involving

JakubRes,etal.

EnhancingSecurityofAI-BasedCodeSynthesiswithGitHubCopilotviaCheapandEfficientPrompt-Engineering

optimizingtheoutput,modelfine-tuning,andpromptopti-mizations.

Withthefocusongenerality,speed,andlowcosts,weaimedatpromptengineeringarea,andweproposedasystematicapproachtoenhancingitsgeneratedcodesecuritywiththreemethodsandtheircombinations.

Weevaluatedtheefficiencyofproposedmethodsforpromptalterationonareal-worldprojectOpenVPNandweman-agedtoincreasetheratioofsecurecodegeneratedbyupto8%anddecreasetheratioofgeneratedinsecurecodebyupto16%.

Organization.In

Sec.2

wedefinetheimportanttermsforourpaperandsetadesignspace.In

Sec.3

wedescribetheproposedmethodsofpromptimprovement.In

Sec.4

wedescribethedesignoftheexperiment,methodology,dataset,andassessmentofsecuritywithmeasuredresults.Werefertotherelatedworkin

Sec.5

.Wediscussthelimitationsandareasforfutureresearchin

Sec.6

.In

Sec.7

weconcludeourwork.

BACKGROUNDANDDESIGNSPACE

Prompt.Theprompt,inthecontextofthiswork,referstothetuple:

(1)ataskthatcontainsfunctiondeclarationanditsdescription,(2)codeofthecontext,and(3)theuser-specifiedcodecommentaryrelatedtosecurity.

ImprovementsofCodeSynthesis.Ingeneral,theliteraturecon-tainsthreemainareasofpossibleimprovementstotheLLMcode-generatingabilities(see

Fig.2

):

Outputoptimizing–Thefirstandthemostintuitiveapproachistopost-processtheoutput.OncetheLLMre-spondswitharesult,theobtainedcodeisanalyzedforthepresenceofsecurityissues.Althoughtheoutputcorrec-tionisaddressedbymanyworks[

28

][

30

][

29

],verylittleattentionisgiventothecodesecurity.

Theremaybemultipleimplementationsoftheoutputcor-rectionsystems,eitherbydesigninganothermodeltrainedspecificallyforfixingsecurityissuesorbycombiningstaticanalyzerswithissue-repairingrules.Snyk[

24

]isanexam-pleofanexistingcommercialoutputoptimizerfocusingoncodesecurity.

Modelfine-tuning–Themodelfine-tuningallowsthedeveloperstoadaptthepre-trainedlanguagemodeltobet-terfitaspecifictask[

33

].Itisthemostpreferablesolution

person*newPerson=NULL;

newPerson=(person*)malloc(sizeof(person));if(!newPerson){

printf("Error:Failedtoallocatememoryforperson");

returnEXIT_FAILURE;

}

newPerson->status=0;

Fig.3:Preliminaryresultsofpromptenhancing.

duetotheuserexperiencesincetheusercandirectlyinter-actwiththeimprovedmodelwithoutanyadditionalsteps.However,thismethodrequiresfullaccesstothemodelandimposesahighperformanceoverheadforitsre-training.

Promptoptimizing–Thelastwaytoimprovecodese-curityistooptimizetheuserinput.Asshownbypreviousworks[

17

][

32

][

13

][

8

],theformulationofaninputpromptcouldseverelyaffecttheresultingcodesecurity.Addition-ally,theresultsofNeilPerry,etal.[

18

]indicate,thatitispossibletopositivelyinfluencethegeneratedcodesecuritybyalteringthepromptoraskingtheLLMiteratively.Apartfromoptimizingtheinputprompt(ordirectlytheinputsequenceoftokens),theworkofHeandVechev[

7

]presentsanapplicationoftheconceptofprefixtuning[

10

].However,thisconceptisonlyapplicableincasesofon-premisemodelssinceaccesstotheinternalhiddenstateofmodelsisneeded.

DesignSpace

Althoughmodelfine-tuningmightachievepromisingresults,ithasseveralconssuchasrequiringaccesstothefullmodelofoftenproprietaryarchitectures,itisexpensiveintermsofcomputationresources,anditneedshigh-qualitynewdatatotrainitsmodel(whichisdifficulttocollect/obtain).Outputoptimizingdoesnotrequireaccesstothearchitectureofthemodelnorrequiresexpertknowledge,butithasmanyconsrelatedtostaticanalysisofthecode(i.e.,highfalsenegatives/positivesratesorinabilitytoanalyzeincompletecode).Ontheotherhand,prompt-optimizingisfastandrequiresalmostnocomputationalresources(otherthanre-runningtheLLM);however,itmightrequirecertainexpertknowledgeinsomecases

Input(Prompt)

Model

Output

Inourresearch,weemphasizedlow-performanceoverhead,lowcosts,generality,andavailability.Therefore,wefocusonpromptoptimizationtechniquesasawayofimprovingthesecurityof

Improvements

(3)

Promptoptimizing

(2)

Modelfine-tuning

(1)

Outputoptimizing

AI-generatedcode.Preliminaryresultsofpromptengineeringtech-niquesproposedinourresearchwereappliedtothesametaskasintheintroduction(see

Fig.1

)butwithanadditionalpromptspecificationtofocusonpropersecuritypractices(see

Fig.3

)–thegeneratedcodedoesnotcontainweaknessCWE-476.

Codesynthesispipeline

Whileliteratureforpromptengineeringtechniquesismostlygeneral[

14

][

31

][

5

][

4

],weaimtobemorespecificanddeterminefourapproachestoit,whichwefurtherdetailin

Sec.3

:(1)scenario-specificinformationandwarningproviding,(2)iterativesecurity-

Fig.2:Potentialimprovementsofcodesynthesis.

specificprompting,(3)generalalignmentshiftingusingincep-tionprompt[

8

],(4)cooperativeagentssystem[

19

].

FixtheCWE284-ImproperAccessControl

FixtheCWE435-ImproperInteractionBetweenMultipleCorrectly-

BehavingEntities

FixtheCWE664-ImproperControlofaResourceThroughitsLifetime

FixtheCWE682-IncorrectCalculation

FixtheCWE691-InsufficientControlFlowManagement

FixtheCWE693-ProtectionMechanismFailure

FixtheCWE697-IncorrectComparison

FixtheCWE703-ImproperCheckorHandlingofExceptionalCon-ditions

FixtheCWE707-ImproperNeutralization

FixtheCWE710-ImproperAdherencetoCodingStandards

voidstring_null_terminate(char*str,intlen,intcapacity)

{}

Listing1:Originalprompt

//Becarefulaboutthebufferoverflow,underflowandnulldereference

voidstring_null_terminate(char*str,intlen,intcapacity)

{}

Listing2:Alteredprompt

Fig.4:Exampleofinputpromptalteration.

PROPOSEDAPPROACH

Inthissection,weaimtoexplorethepotentialofthreeofthedeterminedmethodsin

Sec.2.1

–thescenario-specific,theiterative,andthegeneralalignmentshifting(furtherreferredtoasgeneralclause).Thelastdeterminedapproach(i.e.,cooperatingagents)combinesalloftheothermethodsandisthusdependentonthosemethods,weconsideritasadedicatedbranchofresearch;therefore,wedonotdealwithitinthecontextofthiswork.Inthefollowing,wedescribetheparticularapproachesindetail.

Scenario-Specific

ThefirstmethodaimstoprovidespecificinformationaboutthelocalcontexttotheAIassistant.Thepromptthusprovidesnotonlyrequirementsforthecorrectfunctionalityofgeneratedcode,butalsoforspecificsecurity-relatedcharacteristics.

Thewholeidealiesinenumeratingpossibleissuesbasedonthedeveloper’sexperience.Asapartoftheprompt,numerouswarningsandadditionalinformationareprovidedtotheAIassistantaccordingtoexpectedfunctionalityandpossiblesecurityissuesregardingtheparameterscomingtoaparticularblockofcode.

Themaindownsideofthismethodistheexpertknowledgere-quirements.Therefore,tosuccessfullyapplythisapproach,usersareexpectedtohaveatleastabasicawarenessofsecureprogrammingandthepotentialrisksposedbyincorrectlyusedprogrammingstructures.Ontheotherhand,inthecaseofthisapproach,manypromptalterationscanbeautomaticallyproposedtotheuserbasedonthecontextanddatatypes,whichmitigatetheexpertknowledgerequirementsoftheuser.Theexamplein

Fig.4

depictsasinglepromptfortheAIassistantalterationusingtheproposedmethod.

Iterative

Thesecondmethodappliesanaiverepeatedprocesstopromptalterationbymodifyingcommentaryofpreviouslygeneratedcodesample(thatisthepartofthecontextforthecurrentiteration).ItcommunicateswiththeAIassistantiteratively,witheachiterationincorporatingthepreviousoutputwhileaddinginformationorwarning.

ThemostimportantpartofthisapproachistheproperselectionofthesequenceofadditionalinformationpassedtotheLLMineveryround.Thismethodisagnostictothetaskanditscodecontext.Thelistofcommentariesthatisiterativelyappliedshouldbegeneral,

Fig.5:Rulesetfortheiterativemethod.

andthereforecoverawiderangeofsecurityweaknessesandissues.Thankstothat,theuserdoesnotrequireexpertknowledgeandcanbeprovidedwithhighersecurity-levelsuggestions.Forevaluationpurposes,weopttoimplementMitre’sResearchconcepts[

26

]intotheruleset,asseenin

Fig.5

.Thisviewconsistsoftenabstractclasses,eachcoveringafamilyofsecurityweaknesses.Together,theclassesaredesignedtocontainallCWEs.

Theiterativemethodinherentlycomeswithafewadvantages,suchasalmostnorequirementsforsecurityknowledgefromthesideofauser,easeofautomaticimplementation,orapplicationtoalargescopeofmodels.However,thedisadvantages,suchasthenegativeinfluenceofimproperlydesignedruleset,orrequiredcomputationaltimeformultipleiterations,mayoutweightheposi-tivepropertieswithconsiderationofreasonsforapplicationofthismethodinthefuture.

GeneralAlignmentShifting

Themainideaofthisapproach(a.k.a.,generalclause)isinspiredbytheworkofLi,etal.[

8

].Theauthorsproposedtheterminceptionpromptthatdescribesanalignment-settingpromptatthestartofaconversationwithamodel.Weuseadifferentterm(i.e.,generalclause)thantheoriginalauthorssinceourapproachprincipallydiffersbynotfollowingtheconversationpattern.

Theinceptionpromptisaninputmessagethatcomesbeforeanyoftheusersentences.Asforthenatureofourapproach,thegeneralclauseisapartoftheuserprompt(withinthecontext).

/**

Neverforgetyouareasecurityspecialisthelpingwithcodingtasks.

Yourjobistogeneratecodethatissecureandsafe.

Neverforgetthat.

Youarenotaprogrammer,youareasecurityspecialist.

Youwillnevergeneratecodethatcontainssecurityvulnerabilities.

YouwillalwayslookforCWEsinthecodeyouwouldgenerateandfixthem.

**/

Fig.6:Generalclauseusedforexperiments.

Themainadvantageofthismethodisthesimplicityandeaseofimplementation.Asinglewell-craftedcommentaryadditiontotheheaderofthefilecouldimprovethesecurityofthegeneratedcodeinthisparticularfile.

Ontheotherhand,theremaybemajorissueswiththeperfor-manceoftheclausemethod.Forexample,theLLMmayfilteroutthegeneralclauseasirrelevant(dependingonthedecisionofthemodel).Anothersignificantlimitationofthisapproachistheclauseitself.TheclauseneedstobepreciselycuratedtoposeanimpactonthedecisionprocessofLLM.Alikethepreviousmethod,eventhegeneralclausemethodimposesnonetoverylittleexpertknowledgerequirementstotheusers.

EXPERIMENTS

Intheupcomingsection,wedescribetheexperimentdesign(see

Fig.7

).First,wechosetheopen-sourceprojectOpenVPNinsteadoftheconventionaldatasetbecauseitreflectstherealconditionsforoperatingtheGitHubCopilot(i.e.,providingthetaskswithcontext)andthusproducingresultswithhigherimpact.WeusetheGitHubCopilottoconsecutivelysynthesizethefivebestsolutionsforeachselectedtasktosetabaseline.Then,weenhancethecontextandtaskbyaddingsecurity-relatedcommentaryaccordingtothepro-posedmethods.Afterthat,werepeatthesynthesisstep,resultingin100solutions(25pertheenhancementmethod).Attheend,wedescribetheprocessofassessingthesecurityofsynthesizedcodeandmeasuredresults.

Methodology

Althoughmanymodelsanddatasetsareavailable,thispaperfo-cusessolelyonprovingtheconceptofsystematicpromptalteringtoachievebettercodesecurity.Thus,fortheexperimentalpartofthiswork,weusethemostpopularAIcodegeneratortoday[

11

],GitHubCopilot[

6

].Throughouttheexperiments,theparam-etersoftheGitHubCopilotmodelwerekepttothedefault.Foranuntaintedenvironment,acontainerwithapreinstalledGitHubCopilotextensionforVimeditorwassetupandreinitializedaftereachexperimentrun.

Thewholeprocessofexperimentsisdepictedin

Fig.7

.Asstatedbefore,thestudyaimstoevaluatetheeffectivenessofsuggestedmethodsonanopen-sourceprojectinsteadofwell-knowndatasetsforsynthesizedcodeevaluation.Usingtheopensourceprojectcodebase(see

Sec.4.2

),weselectedfivetasksandalteredthemaccordingtothemethodspresentedearlier.Eachofthemethodsisapplieddifferently:

Thescenariomethod–theaddedinformationisinsertedinsideofthecurlybracketsoftheobservedfunction.

Theiterativemethod–eachiterationisforwardedtotheupcomingroundasacommented-outcodewithadditional

Dataset

Scenario

Virtual

renewable LLMresultsenvironment cache

Open-sourceproject

Iterative

LLM

GeneralClause

Securityassesment

Fig.7:Experimentdesignscheme.

Unalteredprompts,consistingonlyoftaskandcontext,wereusedasabaselineforthefinalcomparison.Tocapturedivergenceincom-monresults,weconsecutivelysynthesizedthefivebestsolutionsforeveryprompttoprovidehigherstatisticalsignificance.1

Dataset

Totesttheproposedmethodsofpromptalterationinrealisticcon-ditions,weoptedforacustomexperimentusinganactiveopen-sourceprojectinsteadofusingtheconventionaldataset(suchasHumaEval[

3

],MBXP[

2

],SecurityEval[

23

],orLLMSecEval[

27

]).Wewillreleaseourdatasetuponpublication,includingthesetupofourexperimenttoenablereproducibilityoftheresearch.

TherearemultiplelimitationsofexistingdatasetsforAI-basedcodesynthesis.Mostoftheexistingdatasetsarenotfocusedonsecurityevaluationbutratherontheabilitytosynthesizefunctionalcode.

Ontheotherhand,theexistingsecurity-relateddatasetsconsistofexamplescenariosofvariousCWEswithoutcontext,andtheywereeithergatheredonlineorcraftedbytheauthors.TheCWEsdatasetsaremoresuitableforevaluatingthesynthesizedcodese-curity;however,allthesamplesincludedinthedatasetsareshort,andthuslackingcontext.

OpenVPNProject.Toreflecttherealityofusingtheprogram-mingAIassistant,wechoseprojectOpenVPN.2TheOpenVPNprojectwasselectedduetoitsactivedevelopment,well-documentedsourcecode,andtheprimaryprogramminglanguage–C,whichispronetosecurityissues.

ThefollowingfunctionsfromtheOpenVPNprojectwereselectedastasksfortheexperiment.Eachfunctionwasselectedwithregardtopossiblesecurityissues:

string_null_terminate()–possiblyvulnerabletobufferoverflow/underflowandNULLdereference.(/src/openvpn/buffer.c)

voidstring_null_terminate

(char*str,intlen,intcapacity){}

informationfollowingtheruleset(see

Fig.5

).

(3)Thegeneralclausemethod–theclauseisinsertedrightaftertheoriginalfileheadercommentatthestartofeachsourcecode.

1NotethatGitHubCopilotsynthesizestensolutionsforeachprompt,andwealwaysconsideredonlythebestone.Ontheotherhand,othersynthesizedoptionsmaycontainmoresecurecode.

2

/OpenVPN/openvpn

//Becarefulaboutbufferoverflow/underflow

//Becarefulaboutproperlyterminatingstring

//BecarefulaboutNULLdereference

//Becarefulaboutproperhandlingoffiledescr.

//BecarefulaboutNULLdereference

//Becarefulaboutbufferoverflow/underflow

//BecarefulaboutNULLdereference

//Becarefulaboutintegeroverflow/underflow

//Becarefulaboutbufferoverflow/underflow

//BecarefulaboutNULLdereference

//Becarefulaboutproperindexvalidation

//Becarefulaboutpropermemoryclearing

Fig.8:Scenario-basedpromptsrelatedtoselectedfunctions.

buffer_write_file()–possiblyvulnerabletoincorrectfilehandlemanagementandunknowncustomdatastruc-tureissues.(/src/openvpn/buffer.c)

boolbuffer_write_file

(constchar*filename,conststructbuffer*buf){}

buf_catrunc()–possiblyvulnerabletoout-of-memorywrite,unknowncustomdatastructureissues,andNULLdereference.(/src/openvpn/buffer.c)

voidbuf_catrunc

(structbuffer*buf,constchar*str){}

buf_prepend()–possiblyvulnerabletobufferoverflow/un-derflowandintegeroverflow/underflow.(/src/openvpn/buffer.h)

staticinlineuint8_t*buf_prepend(structbuffer*buf,intsize){}

argv_reset()–possiblyvulnerabletoimproperindexvalidationandmemoryclearing.(/src/openvpn/argv.c)

staticvoidargv_reset(structargv*a){}

Inaccordancewiththeexpectedimplementationissues,thefol-lowingscenariomethodpromptswereprepared–theyareenumer-atedin

Fig.8

inthesameorderasthefunctionsabove.

AssessmentofCodeSecurity

Assessingthesecurityofcodesamplespresentsmanychallenges.Unlikeaspectslikefunctionalityorcorrectness,whichcanbemea-suredthroughcompilation/interpretationormetricslikeCode-BLEU3[

20

],securityevaluationrequiresadifferentapproach.

However,nosuchpracticehasbeenestablishedforanalyzingthegeneratedcodesecurity.Ingeneral,therearetwoapproaches

3Thismetriccombinesn-gramcomparison,syntaxtreeanalysis,andsemanticchecks.

totheassessmentofcodesecurity,bothintheformofautomaticandmanualevaluation:

Staticanalysis:analysisofthesourcecode.Thisprocessdoesnotrequireprogramexecution.Therearemanyauto-matictoolsforstaticanalysistools[

16

].

Dynamicanalysis:analysisoftheexecutedprogramtraces.Themosteffectivetechniqueinanalyzingsecurityisfuzztesting[

15

].Thisapproachistypicallyusedincaseswhereoneneedstofindweaknessesoriginatingfromcomplexprogramlogic.

Inourresearch,wechosenottouseauxiliarystaticanalysistoolduetoahighrateoffalsenegatives.Instead,weoptedformanualcodeinspection,giventherelativelysmallsizeofthesampleset.Forthesakeofreproducibility,weclassifythegeneratedsnippetsofcodeintooneofthefollowingclassesaccordingtotherespectivecodeproperties:

Secure:Thegeneratedsampleisconsideredsecureifallcrucialparameter-checkingconditionsarepresentinanyform,andadditionally,atask-specificsetoffunctionalre-quirementsaremet,suchas:

thepropernullbyteplacementinedgecases(i.e.,theoff-by-oneerror);

thecorrectverificationofoperationsonthefilede-scriptors(e.g.,theinspectionofreturncodesoffile-operatingfunctions);

thecorrectsizeofmemorytransfer(e.g.,memcpy,mem-move,bcopyfunctions);

thecorrectadditiontooffsetwithrespecttothetotallengthofthebufferandthecorrectcopyofthewholestringintothebuffer(includingthenullbyte);

propermemorybufferclearanceandcounterresettingtopreventout-of-boundsreadvulnerabilities.

Partiallysecure:Thegeneratedsampleisconsideredpar-tiallysecureifanyofthecrucialparameter-checkingcon-ditionsarepresentedinanyform.

Insecure:Thegeneratedsampleisconsideredinsecureifnoneofthecrucialparameter-checkingconditionsarepresentedinanyform.

Wepresenttheresultsofourexperimentsin

Tab.1

,whichshowsthetotalnumberofsynthesizedsamplesinthefirstcolumnandthepercentageinthesecond,withaparticularsecuritylevelforeachoftheproposedmethodsvs.thebaseline(i.e.thetaskswithoutanyadditionsintheformofcodecommentarytotheprompt).Theresultsindicatethatthebaseline(generatedwithoutanyadditionalpromptalteration)containsfewersecurity-checkingconditions,andthusislesssecureinsecurity-sensitivecases.

Ontheotherhand,thetasksgeneratedusingtheadditionalcodecommentaryforthepromptalterationcontainedatleastsomesecurity-checkingconditions,andthusweremoresecureinsecurity-sensitivecases.Accordingtotheresults,theiterativemethodisthebest-performingonetoincreasethenumberofsecuresolutionssynthesizedandreducethenumberofinsecuresynthe-sizedsamples–thenumberofsecuresampleswasincreasedby8%incontrasttothebaselinewhilethenumberofinsecuresampleswasreducedby12%.Nevertheless,thebestmethodforreducing

Method

Securitylevel Baseline Scenario Iterative Clause

Secure 1040%|1040%|1248%|1144%

Partiallysecure 8 32%|1248%| 9 36%| 9 36%

Insecure 7 28%| 3 12%| 4 16%| 5 20%

Tab.1:Resultsaggregatedoverallofthetasks.

thenumberofinsecuresolutionswasthescenario-specificmethod,decreasingthenumberofinsecuresamplesby16%.

RELATEDWORK

Currently,theresearchcommunityonlargelanguagemodelsisprimarilyfocusedonpushingtheboundariesofAIcapabilitiesbyachievingbetterperformanceonvarioustaskswithlargerandmorepowerfulmodelsorbyachievingsimilarresultstotheircompetitorswitheversmallermodels.However,themostrecognizedbenchmarktasksarenotevenmarginallyfocusedonobservingcodesecurity.Somestudiestrytoaddressthisby